Central processing units are emerging as the next beneficiary of the artificial intelligence boom, driven by the strong debut of Meta's personal AI agent "Muse." While AI investment has centered on graphics processing units, expectations are growing that CPUs will take on a larger role as AI evolves from chatbots that answer questions into agents that carry out tasks on behalf of users. In South Korea, exchange-traded funds tied to CPUs have topped the performance rankings.
CPU-related products dominated the top of the domestic ETF return rankings over the week from Sept. 16 to 23, according to ETF CHECK data released on the 24th. KODEX US CPU Semiconductor TOP10 rose 22.65% for the biggest gain overall, while KIWOOM US CPU Semiconductor TOP4+ climbed 16.18% for third place. ACE Global AI Custom Semiconductor ranked second with a 19.39% gain.
KODEX US CPU Semiconductor TOP10 concentrates its holdings in leading CPU makers. As of the 23rd, Intel accounted for the largest share at 26.59%, followed by AMD at 25.96% and Arm at 24.94%. The top three holdings alone made up more than 77% of the fund. It also holds Qualcomm at 6.51% and Marvell Technology at 6.43%, meaning the recent rally in CPU-related shares fed directly into the ETF's returns.
Behind the rally is Muse, the personal AI agent Meta launched on the 8th of this month. Unlike chatbots that simply answer user questions, Muse performs actual tasks such as sending emails and booking travel. It is built as a resident system that continues working in the background even after the app is closed.
Early reception has been strong. Muse recorded 1.8 million cumulative downloads within 12 days of launch, surpassing the 1.3 million that ChatGPT logged over the same period after its release, according to Eugene Investment & Securities. Daily active users reached 640,000, far above the 230,000 ChatGPT had at the comparable point.
What the market is watching is the increase in CPU demand as the center of gravity in AI services shifts from chatbots to agents. AI agents repeatedly call external programs, access databases and run fresh rounds of inference. While GPUs handle AI computation, CPUs play the central role in linking and coordinating each of those tasks.
Muse in particular assigns a dedicated cloud virtual machine to each user, including those on the free tier, suggesting that CPU demand could rise alongside user growth. Intel has projected that the ratio of GPUs to CPUs could shift from the current level of between eight-to-one and four-to-one to one-to-one or higher over the long term. AMD, for its part, has put the server CPU market at $220 billion by 2030.
The CPU boom is spreading beyond chipmakers to South Korean semiconductor component and substrate suppliers. If the spread of AI agents lifts the volume of server CPUs deployed, demand should also rise for multilayer ceramic capacitors, Ajinomoto build-up film semiconductor package substrates and high-end printed circuit boards used in CPUs.
Meritz Securities said it had confirmed that Samsung Electro-Mechanics (009150.KS) signed a long-term MLCC supply contract with a North American CPU maker. Shipments of server-grade ABF substrates to a North American CPU maker are also rising at the company's Vietnam plant, and the brokerage estimated that CPU-bound products account for more than 45% of total ABF revenue when PC applications are included. Expecting stronger component demand from an expanding AI market to translate into earnings growth, it maintained a buy rating and raised its target price to 2.2 million won from 2.1 million won.
Expectations are also building for growth in the market for high-end PCBs used in AI servers. Shinhan Securities pointed to rising demand for high-layer-count, highly integrated substrates at Isu Petasys (007660.KS), along with the spread of multilayer stacking and high-density interconnect boards, as growth drivers. Estimating that the average selling price of the company's PCBs rose 21% from a year earlier in the second quarter, the brokerage set a target price of 170,000 won, 6% above its previous level.
"The faster agent services spread, the more likely it is that demand for CPUs handling workload orchestration will expand structurally," said Park Jae-hwan, an analyst at Eugene Investment & Securities. "The already tight bottleneck in CPU supply is likely to intensify further."
